Nature Neuroscience

Conformal prediction is increasingly proposed for clinical decision support because it provides distribution-free coverage guarantees that hold for any model and any data distribution. The guarantee routinely reported, however, is marginal, and we show that in multi-class diagnostic classification it can be satisfied exactly while individual diagnoses are covered very unevenly. On a four-way mood and psychosis classification task over 1,520 subjects from three studies and 14 acquisition sites, marginal split-conformal calibration attained 0.9000 empirical coverage against a nominal 0.90, while healthy controls were covered at 0.941 and schizoaffective disorder at 0.818. The surplus and the deficit cancel, so the aggregate figure is uninformative about either. Class-conditional (Mondrian) calibration reduced this 12.3-point disparity to 0.4 points at a cost of 0.07 labels in mean set size, under 3%. Holding coverage fixed across diagnoses has a further consequence: the residual variation in set size can no longer be attributed to class prior or per-class accuracy, and becomes interpretable as a property of the subject. We show that the resulting prediction sets separate subjects into confident, boundary, ambiguous and unresolved strata, and that the proportion of subjects independently flagged as label-ambiguous by a structural-MRI model trained separately on the same cohort rises monotonically across these strata (34.1%, 57.3%, 68.1%, 81.8%; p = 8.8e-18). No schizoaffective subject and 0.9% of bipolar subjects reach the confident stratum, against 18.0% of controls and 13.4% of schizophrenia subjects. Set size at matched coverage also provides a comparison between representations that accuracy cannot. Applied to structural MRI, functional MRI and their fusion, it reveals a sign reversal that accuracy conceals: fusion reduces set size for bipolar (-0.25) and schizoaffective (-0.23) subjects and increases it for controls (+0.16), and the aggregate fusion benefit itself changes sign below alpha = 0.10, while top-1 accuracy rises monotonically from 0.495 to 0.578 to 0.611 across the three models. Reporting marginal coverage alone is insufficient whenever diagnostic classes are unbalanced.

Chronic pain mechanisms are complex, spanning multiple brain regions and networks. We ask whether resting brain activity carries a readout of that state. From a few minutes of resting-state electroencephalography (EEG), we generate a spectrogram to represent how each region of the cortex oscillates across frequency and time and pass it through CREST (Cortical Resting-state EEG Spatial Transformer): a frozen image-recognition network that reads each region as an image--here, a spectrogram--paired with a graph model that weighs the 56 cortical regions together to classify chronic-pain status. Across 125 people (74 with chronic pain, 51 healthy controls), evaluated through a leave-one-subject-out cross-validation, CREST separates the two groups with an area under the receiver operating characteristic curve (AUROC) = 0.782 (permutation p < 0.005). Control experiments implicate each persons individual alpha rhythm. Clinical relevanceA resting-state EEG readout of chronic MSK pain could clarify pathophysiology and inform treatment.

The latent cause theory of memory modification provides a computational account of how the brain decides whether to update existing memories or form new ones, but leaves unspecified the neural mechanisms implementing this inference. We propose Reuse-and-Append Memory (RAM), a set of mechanistic principles that achieve the same goal through sparse neural coding, Hebbian learning, and pattern-matching dynamics. The core idea is that neurons encoding prior experiences are automatically reactivated by similar stimuli, while uncommitted neurons are recruited to encode genuinely novel aspects of each experience, including the passage of time. We present a computational model instantiating these principles and show that it reproduces acquisition, extinction, renewal, and spontaneous recovery in fear conditioning. By committing to a neural mechanism, we found that RAM separates what is typically modeled as a single prediction error driving memory updating or formation into two independent signals: an immediate novelty signal that emerges from coverage-based allocation of new neurons, and an outcome mismatch signal that later activates safety circuits when an expected outcome fails to arrive. RAM also reproduces the dependence of recovery on the timing of stimulus reminders (the Monfils-Schiller effect), but predicts it to be inherently fragile, consistent with the mixed empirical record. RAM suggests that the Monfils-Schiller effect, often attributed to reconsolidation, leaves the original fear memory intact: a bridging event carries safety information to new contexts through offline co-retrieval. More broadly, none of the phenomena we explain require modifying existing memories, and the most stable configuration is the extreme form of the principle: always append, never overwrite.

How sensory cortical activity is read out by downstream circuits to guide behavior is a fundamental unsolved problem. Substantial work has examined correlations between sensory neural responses and animals perceptual judgments, but their interpretations remained controversial due to many intervening variables, such as responses of unrecorded neurons. Furthermore, stimulus and choice encoding in sensory populations are often not well aligned, and it remains contested whether this misalignment indicates a limitation in sensory readout. Here, we introduce a closed-loop, neurofeedback paradigm that directly interrogates the capacity of sensory readout: the key idea is to train subjects to report specific patterns of population activity in a sensory area recorded online, rather than the actual stimuli presented. To test this, we trained macaque monkeys on a visual change-detection task using shape stimuli, implanted an electrode array in visual area V4, and tested whether they could be further trained to rely on their own V4 activity along specific axes in neural state space. Strikingly, monkeys successfully increased the neuron-choice correlation along trained axes in neural population state space. No detectable changes in stimulus selectivity or noise correlations were found within the recorded population, and further model simulations confirmed that adjustment of sensory readouts best accounted for the results. A control experiment that merely disrupted the stimulus-reward contingency without a closed loop failed to enhance neuron-choice correlation. Together, these results demonstrate that closed-loop neural feedback achieves neuron-choice alignment beyond the ceiling of natural perceptual training, suggesting that the misalignment in perceptual tasks reflects constraints on learning within naturally available training regimes.

Understanding how the brain supports flexible, goal-directed behavior is a central aim of neuroscience. Yet such behavior is highly variable and unfolds over long timescales, making it challenging to understand its organization. Advances in behavioral tracking quantify movements but do not reveal the objectives that guide them. Motivated by this, we develop inverse reinforcement learning (IRL) into a general framework for inferring these objectives from naturalistic behavior. Our approach represents an animal's objectives as internal rewards, and infers these rewards as well as the policies they induce from observed behavior. It applies across diverse behavioral settings, from freely moving behaviors to structured tasks. It can infer objectives that switch over time or gradually change with experience. In mice capturing crickets, it identified a small set of objectives that the animals switched between over the course of a trial. In mice learning a cued labyrinth, it inferred how the animals' internal rewards changed as learning progressed. Across both datasets, the models predicted held-out actions and generated extended behavior that reproduced key features of animal performance. By transforming long, variable behavioral sequences into time-resolved estimates of the animal's objectives, our approach provides a quantitative basis for studying how the brain organizes natural behavior.

Acquired central nervous system (CNS) injury is one of the most common neurological conditions globally, yet effective treatment options are lacking. As the main tissue-resident macrophages of the CNS, microglia have emerged as key functional regulators of CNS repair. However, means to induce neuroprotective microglia and harness their intrinsic repair capabilities have remained elusive. Here we identify gp130 as a key receptor molecule for facilitating bidirectional microglia-neuron communication that improves outcomes from CNS injury. We show that activation of gp130 in CNS-resident microglia triggers the secretion of leukemia inhibitory factor (LIF), a neurotrophic cytokine. LIF induces neuronal IL-6 secretion that then acts back onto the microglial gp130 receptor, thus creating a neuroprotective loop. We demonstrate the broad therapeutic potential of acute gp130 activation across multiple models, including traumatic brain injury, stroke, and spinal cord injury, and that this pathway can be leveraged therapeutically with designer cytokines.

Adaptive behavior requires external input to be interpreted together with internal state and ongoing behavior. During pain, noxious somatosensory input evokes movement and arousal, and injury reshapes this relationship, yet how cortical activity organizes stimulus content with behavioral state remains unclear. In awake mice, we delivered hindpaw stimuli while tracking movement, arousal and facial expression, and studied primary somatosensory cortex (S1) using widefield and two-photon calcium imaging, single-action-potential activation of nociceptors, and S1 silencing. Here, we show that S1 neurons were broadly recruited by stimulus and state, whereas latent population dimensions carried mechanical stimulus content. Inflammatory injury caused a reorganization of S1 geometry, binding state and protective responses tighter together. Noxious heat drove S1 as strongly, but engaged mostly the state axis. S1 silencing reduced mechanical hypersensitivity, arousal, and facial expressions. S1 thus embeds mechanical input within a stimulus-state geometry, which is reorganized during inflammatory injury to support adaptation of a coordinated protective response.

For datasets with thousands of neurons and images, vision transformers have proven successful at predicting neural responses to stimuli. However, they are expected to underperform in low-data regimes, where CNNs and Gaussian processes are considered more effective. We ask whether transformers can be made competitive for small-scale neural prediction, and show that underperformance in this regime can be overturned with the right inductive bias. We equip a vision transformer with a differentiable per-neuron circular crop in feature space. The crop is centered on each neuron's receptive field, with a radius selected per neuron, so the model only sees the small image region that drives that neuron instead of the whole image. This makes the cost of attention scale with the size of the receptive field rather than with the size of the image. A single transformer stack is shared across all neurons: each neuron's specificity resides in the crop, not in the architecture. We call this model circular receptive-field vision transformer (CiRF-ViT). We evaluate it on small multi-electrode-array recordings of mouse and salamander retinas: a mouse preparation of 41 ganglion cells and two salamander preparations totaling 49 ganglion cells, each with only a few thousand stimulus-response pairs, two orders of magnitude below the scale at which transformers are typically trained. Against CNN and Gaussian-process baselines, CiRF-ViT reaches the highest mean explained variance on both datasets (0.94 on mouse, 0.95 on salamander). Probed with the local spike-triggered average (LSTA), a zero-shot test of context-dependent, nonlinear stimulus sensitivity, CiRF-ViT reproduces the qualitative polarity inversion that a linear model cannot capture by construction. A biologically grounded, per-neuron locality prior is therefore enough to make transformers competitive for neural prediction well below their usual data scale, while matching the reference models on an established functional signature of the retina's nonlinear computation.

Neural recordings from different individuals vary substantially even when behavior is broadly shared. The sampled neurons differ, and the same behavior occurs at different times. Standard cross-subject analyses rely on matched time points or anatomical correspondence, which excludes many datasets. We recently introduced Shared Representation Discovery (ShaReD), which identifies neural-behavioral relationships conserved across subjects by learning a shared behavioral projection together with subject-specific neural projections. Here we develop and benchmark this method using synthetic, primate, and rat data. On synthetic data, ShaReD recovers common structure across noise levels, sample sizes, and subject counts, and separates components confined to different groups of subjects. In non-human primate motor cortex, ShaReD identifies kinematic representations that generalize across individuals and across reaching tasks with different movement statistics. In rats navigating a spatial alternation task, ShaReD isolates behavior-aligned directions within the CA1-to-PFC communication subspace. ShaReD thus extends multi-subject analysis to datasets in which comparable behaviors occur without cross-subject temporal correspondence.

Perception and action show tight links to the statistical structure of physical stimuli and likely rewards, but the underlying mechanisms are unknown. A simple, biologically plausible network model shows that probabilistic behavior emerges naturally in diverse scenarios, and arises from sampling of competing responses. Notably, it also provides a principled computational rationale for the prevalent finding of balanced excitation and inhibition in the brain. Recurrent connections within a population of excitatory neurons embed multiple attractor states, and coupling to a pool of inhibitory neurons enforces mutual exclusivity among the states. Upon concurrent stimulation, competing attractors alternate in activity. These global state transitions are caused by local, uncorrelated spiking noise and yet convey, over time, the relative strengths of attractors' support. The simplest probabilistic competitive recurrent networks (PCRNs) allow for closed-form analysis, shedding light on the neural basis of choice behavior under uncertainty. More complex systems of laterally connected PCRNs can collectively resolve the myriad local ambiguities pervasive in sensory stimuli, rapidly settling into globally-consistent configurations that match perceptual reports. Alternations are crucial in all cases, and occur only if a PCRN's inhibitory pool is strong enough to prevent attractors' activity from reaching saturation. A balance between excitation and inhibition is thus both a prerequisite and a hallmark of probabilistic sampling in cortical networks.

Whether the heterogeneity of psychiatric conditions converges on shared neurophysiological alterations or translates into distinct signatures remains unclear. We assembled high-density electroencephalogram (hd-EEG) resting-state recordings from 4,812 individuals aged 5 months to 66 years across 11 psychiatric conditions, a broad spectrum of rare genetic variants, and typically developing (TD) individuals. We established normative developmental trajectories of source-space EEG across spectral organization, connectivity, and signal complexity. Psychiatric conditions showed small deviations, revealing a shared transdiagnostic profile. In contrast, single rare variants showed substantially larger, distinct and sometimes mirror-opposite signatures that collapsed toward the psychiatric profile when pooled. Autism Spectrum Disorder showed some of the smallest group-level effects yet the largest individual deviations, indicating substantial but directionally inconsistent alterations. EEG deviations followed a cortical gradient, with larger effects in sensorimotor regions. We demonstrate that sample sizes in the hundreds are required for robust associations with psychiatric diagnoses. This interactive open resource provides normative scores to benchmark future results.

Neural populations in auditory cortex (AC) perform sensory computations that support stimulus category decoding and flexible behavior. The underlying geometry and generalizability of these computations for natural stimuli remain poorly understood, particularly at the single-neuron level, where neurons display wide-ranging tuning specificity and temporal acuity. To address this gap, we developed ACNet, a foundation model of cortical sound encoding, to predict the time-varying activity of >3000 neurons in AC of ferrets. Training data included 42 hours of natural sounds spanning over 100 categories and were collected from multiple recording sites across multiple animals. The model achieved state-of-the-art response prediction accuracy. Model activity was succinctly captured by a low-dimensional neural manifold, which generalized (>80% of variance) across animals. Analysis of ACNet activations revealed the emergence of rate-based, sparse sound coding across layers, a prominent feature of the auditory cortex. These transformations were concomitant with the emergence of more accurate and neurally aligned auditory category decoding, even though ACNet was not explicitly trained to categorize sounds. The model also revealed tuning differences between anatomically distinct cell types. Taken together, our results demonstrate that foundation models of sensory systems can reveal generalizable computations by large neural populations.

Clinical improvement following psychiatric neuromodulation often requires weeks to months, creating prolonged uncertainty during treatment optimization. Neural biomarkers could reduce this uncertainty by forecasting treatment outcome, providing early evidence of therapeutic engagement, and objectively tracking clinical response. We analyzed nearly 120,000 patient-hours of chronic intracranial recordings from 24 patients undergoing ventral capsule/ventral striatum deep brain stimulation for treatment-resistant obsessive-compulsive disorder. Before stimulation initiation, circadian neural features distinguished eventual responders from non-responders. Within the first two weeks of therapy, reductions in neural predictability differentiated eventual responders from non-responders and correlated with the magnitude of subsequent symptom improvement, months before stable clinical response was achieved. During chronic treatment, neural predictability tracked clinical response state. Together, these findings establish a framework for neural biomarkers spanning the course of DBS therapy: forecasting treatment response before stimulation, detecting early therapeutic engagement after activation, and longitudinally monitoring clinical response.

Single-cell RNA sequencing (scRNA-seq) captures detailed gene expression profiles at scale, while patch-clamp recordings measure intrinsic neuronal electrophysiological properties. Modeling the relations between these two modalities remains a challenge. Here, we compare how well electrophysiological features can be predicted by traditional transcriptomic cell type classification, representations derived from a foundational model (scGPT) pretrained on large-scale scRNA-seq datasets, ion channel-coding genes, and highly variable genes. Using paired transcriptomic and electrophysiological patch-sequencing data from 495 human neurons from neurosurgical tissue, we find that cluster-level cell type representations consistently outperform highly variable gene selection, ion channel gene selection, and context-enriched scGPT embeddings. Notably, performance varies across model architectures and initializations, and the best results are obtained by combining the outputs of separate cell type and scGPT-based models. Together, these findings suggest that traditional discrete cellular classification is highly effective in predicting physiological features. For maximum performance it can be complemented by pretrained transformer models.

Background: Traumatic brain injury (TBI) pathology evolves over hours to months, and molecular staging could complement clinical and imaging-based assessment. Single-cell transcriptomics captures cell-type-resolved temporal programs, but whether a transcriptional "molecular clock" trained on one injury model can be transferred to others remains unknown. Methods: We compiled four public mouse TBI single-cell/single-nucleus datasets spanning two injury models - controlled cortical impact (GSE277487; CEREBRI) and fluid percussion injury (GSE160763) - plus an independent mild-FPI cohort reserved for validation (GSE247339), totalling 307,991 cells across four pathological stages (control, acute, subacute, chronic). Classifiers were evaluated at cell and sample levels (held-out libraries) against a permutation null, with per-cell-type, strict leave-one-dataset-out (LODO) generalization, glial-only transfer and gene-level SHAP analyses. Results: Within GSE277487, sample-level balanced accuracy reached 0.67 (permutation null 0.32 +/- 0.09; p = 0.001), and the signal was reproduced in the independent mild-FPI cohort (0.40 vs null 0.19 +/- 0.05; p = 0.0010). Cell-level cross-validation overestimated performance by ~0.16, quantifying within-sample leakage. Cross-dataset generalization was poor (strict LODO balanced accuracy 0.20-0.31); the acute immediate-early response was untransferable (recall 0.002) and 80% of chronic (6-month) cells were assigned to the subacute class. Glial-only models raised subacute recall across models (0.91 and 0.75), but precision was indistinguishable from a subacute-only classifier (0.51 vs 0.50; 0.58 vs 0.58) and one-vs-rest AUROC was at or below chance (0.38-0.54) across four classifier families - classifier collapse, not transferable discrimination. The reproducible cross-model component was instead a gene-expression programme: a microglial complement/lysosomal damage programme (Trem2-Apoe axis) was up-regulated across the severe CCI and FPI models (13/13 concordant genes), whereas the immediate-early module was model-specific. Conclusions: A single-cell molecular clock for TBI staging is feasible within datasets once within-sample leakage is controlled, but it does not transfer across datasets as a classifier: apparent cross-model transfer reflected classifier collapse onto the majority class, and the reproducible cross-model component was a microglial complement/lysosomal gene-expression programme. These findings define the boundaries of cross-dataset time inference in TBI and provide evaluation criteria for claims of transferable single-cell classifiers.

Can we tell what animals are dreaming about from their brain activity? This fascinating question has been explored extensively using the rodent hippocampus, where place cells are sequentially reactivated during sleep, consistent with replay of paths explored while awake. However, dreams involve vivid visual experiences, and whether analogous replays occur in the visual cortex has not been explored, that too in primates. To address this gap, we performed wireless neural recordings from the high-level visual (IT) and motor regions (PMv, PFC) in two monkeys while they viewed food images on a screen while awake, and afterwards while they slept in their natural setting. Decoders trained on sleep/wake neural activity in each region were able to accurately predict the ongoing sleep/wake state, as well as upcoming transitions between states. Interestingly, decoders trained to categorize neural responses to food (but not faces/objects) showed increased probability during REM-like sleep compared to NREM or awake states. Thus, behaviorally relevant visual stimuli seen while awake are reactivated during REM-like sleep. In other words, these monkeys saw food, faces and objects while awake, and dreamt selectively about food while sleeping.

Extracellular recordings carry information about neuronal location, and pooling per-unit predictions across neighboring electrodes improves its readout, but whether this local pooling should be learned is untested. We present MYCEL (Message-passing Yields Cellular Embeddings of Location), a graph neural network whose nodes are spike-sorted units carrying waveform and spike-timing features selected on cell-type ground truth, and whose edges encode recording geometry. On the Allen Brain Observatory Visual Coding Neuropixels dataset, learned message passing outperforms fixed aggregation, with the largest gains where single units are least informative. On the International Brain Laboratory Brain Wide Map, MYCEL matches the strongest population model at one-tenth the pairwise interactions. On held-out animals, the embeddings recover laminar depth within visual cortex. Retrained with only the edge geometry changed, the same framework decodes anatomy in macaque visual cortex and human medial temporal lobe. Thus, how neighboring neurons are combined matters as much as the features that describe them.

Regeneration of central nervous system (CNS) axons depends on Transcription Factors (TFs) that reactivate developmental growth programs, yet most such factors remain unknown. By intersecting developmental chromatin binding with pro-growth gene networks, we identified two retinoic acid receptor transcription factors, RARA and RARG, whose occupancy at growth-associated genes is progressively lost as neurons mature. Restoring both factors together increased neurite outgrowth beyond either alone in two independent systems, the Neuro-2a cell line and primary cortical neurons. In vivo, the same combination drove cross-midline sprouting after pyramidotomy and long-tract regeneration after thoracic spinal cord crush, with concordant recovery of hindlimb gait and grip strength. Interestingly, neither receptor alone was sufficient, hinting at combinatorial regulation. Single-nucleus transcriptomics delineated that only the combination reactivated relevant cytoskeletal and gene-expression programs, while genome-wide binding maps showed that RARA and RARG partition the regulatory landscape, with RARG dominating promoters and RARA occupying distal enhancers, so that neither receptor reconstitutes the developmental growth state alone. Intriguingly, this cooperative requirement was specific to the CNS: in peripheral sensory neurons, RARG alone was sufficient and RARA was dispensable. These data identify RARA and RARG as novel cooperative regulators of regenerative axon growth in mammalian CNS and PNS neurons and potential targets for therapeutic intervention.

Leading deep neural network encoding models predict visual cortical responses with nearly indistinguishable accuracy, raising the strong inference that these models have converged on the same underlying brain-aligned parameterization of natural image space. Here we demonstrate that this is not the case. We introduce axis-aligned feature accentuation, which converts each model's fitted encoding axis into graded stimulus perturbations that are predicted to parametrically control neural firing within and beyond the natural-image range. We generated over 27,500 controller stimuli from ten leading vision models and presented them to five macaques in closed-loop experiments targeting early, mid-, and high-level visual areas. Despite matched natural image predictivity, models diverged strongly in their ability to control neural firing using accentuated stimuli, revealing that most model encoding axes failed to capture the precise tuning of their corresponding neurons. The two adversarially trained models showed a consistent advantage, though adversarial robustness was only weakly predictive of neural control across other models. Instead, control was better predicted by the spatial frequency structure of the input gradient: the distribution of pixels influencing each encoding axis. Overall, these results establish neural control via axis-aligned feature accentuation as a causal method to assess the alignment between how neurons and models parameterize the visual world.

To act adaptively, animals must infer hidden states of the world from incomplete sensory information and update beliefs as new observations accrue. While dopamine signals are well explained by reinforcement learning models that incorporate belief states, how the brain implements belief-state inference remains unknown. Prior modeling work showed that recurrent neural networks trained to predict value (Value-RNN) develop task-specific, attractor-like dynamics that mirror evolution of beliefs. Here we performed high-density electrophysiological recordings from orbitofrontal cortex (OFC) and other brain areas of mice performing two variants of a Pavlovian task that differ in reward probability, which produce different within-trial belief dynamics. We find that OFC population activity exhibits task-specific attractor-like dynamics that mirror the within-trial dynamics in the Value-RNN. These dynamics are absent in motor and olfactory regions, are not explained by behavioral differences, and emerge progressively with learning. Our findings indicate that the brain approximates belief-state inference through learned, task-specific attractor-like dynamics.